Adaptive operator selection with bandits for scaled problems based on decomposition-based MOEAs
YunFei Tang, Shuang Li, Wei Li, Gang Luo · Engineering Optimization · 2025
Decomposition-based multi-objective evolutionary algorithms (MOEAs) are widely utilized for solving multi-objective optimization problems (MOPs), while having the problem of striking a balance between exploration and exploitation. To address this issue, many efforts have focused on adaptive operator selection mechanisms that automatically choose the most appropriate operator according to the improvement of the evolutionary population. However, real-world MOPs typically feature objectives with varying scales. Existing operator selection mechanisms struggle with such scaled MOPs, since most, if not all, of the improvement calculation does not consider the scales of objectives. This article introduces a novel adaptive operator selection mechanism designed for scaled problems. Specifically, a multi-bandit learning model is formulated, where each arm represents a reproduction operator and is assigned a prior reward distribution. These rewards are calculated based on the normalized objective values, and the operator (arm) with the highest reward is selected to generate an offspring. Experimental results comprehensively demonstrate the effectiveness and competitiveness of the proposed mechanism on scaled problems, compared with eight well-established MOEAs.